Individual Treatment Effect Prediction Using Model-Based Random Forests
Established statistical procedures for the analysis of primary endpoints in randomized clinical trials assume that there is a universal (i.e., constant) treatment effect that applies to all patients in the trial and - even more importantly - to all future patients potentially to be treated with the novum under consideration. Methods for subgroup analysis relax this assumption and allow the treatment effect to depend on patient characteristics. This means that the treatment effect and thus the potential benefit or harm of a specific therapy is different in subgroups of patients who have been diagnosed with the same disease. We introduce model-based recursive partitioning as a procedure for the automated detection of patient subgroups that are identifiable by patient characteristics. For more individualized treatment effects we propose the use of ensembles of model-based trees, i.e., model-based random forests.